Turbulence super-resolution stops treating density and velocity as separate channels: 4D convolution lifts mean PSNR from 20.39 to 24.47 at 32x coarsening
Synopsis
Three-dimensional compressible turbulence super-resolution has long treated density and the three velocity components as unrelated channels, which leaves standard 3D convolution poorly suited to the coupling between variables. A four-dimensional convolution operator keeps physical variables on an explicit feature axis and shares kernels across the joint spatial-variable domain, and it has been inserted into the EDSR and RCAN super-resolution backbones. On the BLASTNet 2.0 Momentum128 benchmark at 32x coarsening, the EDSR backbone's mean PSNR rises from 20.39 to 24.47, with zero-shot reconstruction on an unseen parametric configuration and on forced homogeneous isotropic turbulence.
Interpretation
Three-dimensional turbulence super-resolution models no longer collapse density and velocity components into unstructured channels; they keep an explicit physical-variable feature axis so the kernel slides along the three spatial axes and this variable axis together. Earlier 3D CNN super-resolution treated different physical fields as ordinary channels and performed only dense channel summation at each layer, without sharing spatial filtering structure along the variable direction. On the BLASTNet 2.0 Momentum128 benchmark at 32x coarsening, the EDSR backbone's mean PSNR rises from 20.39 to 24.47 and mean SSIM from 0.5230 to 0.6214, at 6.0M versus 5.8M parameters.
The construction extends standard convolution to a joint four-dimensional spatial-variable domain, giving the input tensor an axis of variable-feature length alongside the three spatial axes. Four-dimensional convolution is an established mathematical operation; here it is organized into a structured spatial-variable parameterization for compressible-flow super-resolution, paired with 4D channel attention and a PixelShuffle that upsamples only the spatial axes. The operator is integrated into two residual super-resolution backbones, EDSR and RCAN, trained on sub-volumes extracted from 27 DNS configurations and evaluated on the in-distribution test set, a parametric variation set and a forced homogeneous isotropic turbulence set.
Perspective
Evaluation is confined to the single-snapshot four-field reconstruction task of BLASTNet 2.0 Momentum128, where the target fields are density and three velocity components and low-resolution inputs are produced by Favre filtering followed by downsampling; these diagnostics measure statistical and structural fidelity rather than strict enforcement of conservation laws.
The single-snapshot task excludes temporal derivatives, pressure and thermodynamic state variables, so unsteady compressible Navier–Stokes residuals cannot be evaluated directly; experiments follow the benchmark's single-run protocol with a fixed random seed, with no multi-seed error bars or significance tests, and no controlled comparison against factorized 4D filtering, grouped convolutions or pointwise channel mixing.
